Instructions to use Beling86/Ornith-1.5-122B-AutoRound-W3W4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Beling86/Ornith-1.5-122B-AutoRound-W3W4 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Beling86/Ornith-1.5-122B-AutoRound-W3W4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Beling86/Ornith-1.5-122B-AutoRound-W3W4")
model = AutoModelForCausalLM.from_pretrained("Beling86/Ornith-1.5-122B-AutoRound-W3W4", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Beling86/Ornith-1.5-122B-AutoRound-W3W4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Beling86/Ornith-1.5-122B-AutoRound-W3W4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Beling86/Ornith-1.5-122B-AutoRound-W3W4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Beling86/Ornith-1.5-122B-AutoRound-W3W4

SGLang

How to use Beling86/Ornith-1.5-122B-AutoRound-W3W4 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Beling86/Ornith-1.5-122B-AutoRound-W3W4" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Beling86/Ornith-1.5-122B-AutoRound-W3W4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Beling86/Ornith-1.5-122B-AutoRound-W3W4" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Beling86/Ornith-1.5-122B-AutoRound-W3W4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Beling86/Ornith-1.5-122B-AutoRound-W3W4 with Docker Model Runner:
```
docker model run hf.co/Beling86/Ornith-1.5-122B-AutoRound-W3W4
```

Ornith-1.5-122B (AutoRound Mixed-Precision W3W4-50GB)

This repository provides an optimized mixed-precision quantization of ornith-ai/Ornith-1.5-122B created with Intel AutoRound.

The model is quantized specifically to fit within 50.8 GB of GPU VRAM, allowing it to run comfortably on a single 64 GB GPU (such as NVIDIA CMP 170HX / RTX 6000 Ada / A100 80GB / H100) with ample headroom (~13.2 GB) for full 262,144 token context inference.

Quantization Architecture & Precision Breakdown

Subsystem	Layers	Precision	Execution Kernel	Footprint
Outer MoE Blocks	12 Blocks (0–5 & 42–47)	W4A16 (group 128, sym)	Marlin Tensor Cores	~14.4 GB
Middle MoE Blocks	36 Blocks (6–41)	W3A16 (group 128, sym)	Marlin / Humming	~32.4 GB
Attention & DeltaNet	All 48 Blocks	W8A16 (group 128, sym)	Marlin Tensor Cores	~4.5 GB
Routers & Shared Experts	All Blocks	BF16 (Preserved)	Native BF16	~1.2 GB

Active Parameters: ~10 Billion per token (Top-8 routed experts + shared experts) out of 122 Billion total.
Calibrated on: 128 authentic long-context technical documents (4096 tokens each) using AutoRound.
Vision Tower: Stripped for dedicated, high-throughput text causal generation (Qwen3_5MoeForCausalLM).

Serving with vLLM

To serve this model with vLLM:

vllm serve Beling86/Ornith-1.5-122B-AutoRound-W3W4 \
  --dtype bfloat16 \
  --max-model-len 262144 \
  --max-num-batched-tokens 4096 \
  --enable-prefix-caching \
  --enable-chunked-prefill \
  --gpu-memory-utilization 0.95

Downloads last month: 76

Safetensors

Model size

16B params

Tensor type

I32

BF16

F16

Model tree for Beling86/Ornith-1.5-122B-AutoRound-W3W4

Base model

ornith-ai/Ornith-1.5-122B

Quantized

(55)

this model